One-step Latent-free Image Generation with Pixel Mean Flows

This paper introduces "pixel MeanFlow" (pMF), an innovative generative model enabling one-step, latent-free image generation. Diverging from conventional diffusion/flow-based models that employ multi-step sampling and latent spaces, pMF meticulously separates the network's output space from its loss space. This allows for direct image generation from noise in a single evaluation. The model demonstrates impressive performance on ImageNet, achieving strong FID scores at both 256x256 and 512x512 resolutions, thus marking a significant stride towards generating high-fidelity images with improved efficiency for various applications like content creation.